Sequencer-based data path synthesis of regular iterative algorithms

Abstract

In many applications, especially signal processing and matrix computations, algorithms are in a highly regular iterative form; access patterns for most variables are highly regular and uniform. Instead of always storing the values of variables back to and retrieving them from memory or register files, it will be much more efficient and cost effective to let those variables intelligently stay or flow in the data path for future use. In this paper, low cost and simple structured sequencers which are best exemplified by hardware stacks and queues are introduced in the data path for efficiently implementing such a novel concept. Various algorithms are developed to map variables to sequencers and to integrate sequencers into conventional high-level synthesis procedures. Experimental results show very encouraging improvement in the performance of designs as well as significant reduction in hardware cost.